It’s not quite the Skynet that Tesla CEO Elon Musk has warned us about, but researchers from the Musk-based OpenAI initiative have made a breakthrough in AI algorithms using Dota 2 as a testbed. OpenAI’s achievement is remarkable due in part to its scope. Most AI versus human matches — be it go or a computer game — involves a single computer against a single human (as was the case with OpenAI’s victory last year). But OpenAI has managed to train its AI to master competing against humans on a five-player team.
The team of five neural networks that was developed is collectively known as the OpenAI Five and is capable of playing [at an accelerated rate] 180 years with of games during a single day. It takes a lot of firepower to mow down seasoned human players that can collaborate, and the OpenAI Five is no exception. At its core are 128,000 CPUs and 256 NVIDIA Tesla P100 GPUs.
There are an immense number of decisions that must be made to be successful in a complex game like Dota 2, with the AI having to cycle through up to 20,000 moves during the course of a 45-minute match. Compare that to 40 or so moves for chess or 150 for Go (where Google has proved that its DeepMind AI can crush even the most skilled human Go players).
The OpenAI Five began its mastery of Dota 2 initially with self-play as it walked without direction around the map. However, a few hours into the training session, the AI started to tackle more complex tasks such as laning and farming.
“After several days, they consistently adopt basic human strategies: attempt to steal Bounty runes from their opponents, walk to their tier one towers to farm, and rotate heroes around the map to gain lane advantage,” writes the OpenAI team. “And with further training, they become proficient at high-level strategies like 5-hero push.”
OpenAI Five thus far has competed against five human teams, which included a team of OpenAI employees, a team of Valve employees and a semi-pro team. OpenAI Five was able to beat its human opponents in a match on May 15th and then turned around and lost one. However, on June 6th, OpenAI Five was back and “decisively won” all three of its matches. When paired up against the amateur and semi-pro teams, OpenAI Five managed to win out in the first three games against its opponents.
“Our underlying motivation reaches beyond Dota. Real-world AI deployments will need to deal with the challenges raised by Dota which are not reflected in Chess, Go, Atari games, or Mujoco benchmark tasks,” the OpenAI team adds. “Ultimately, we will measure the success of our Dota system in its application to real-world tasks.”